<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-0318">
  <Title>Automatic Disambiguation of Discourse Particles</Title>
  <Section position="3" start_page="107" end_page="109" type="metho">
    <SectionTitle>
2 The Functions of Discourse
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="107" end_page="109" type="sub_section">
      <SectionTitle>
Particles
</SectionTitle>
      <Paragraph position="0"> The following examples from a corpus recorded in the appointment scheduling domain (Verbmobil Database, TP 13) show several different functions the English discourse particle yeah may fulfill:  (1) 13BAR: what about the 18th of December? null 14RIC: &lt;P&gt; yeah, yeah, that work.</Paragraph>
      <Paragraph position="1"> (2) 1UMI: yeah, we've got to get together and discuss &lt;P&gt; Stufe A fiir die Studienordnung, a (3) 124ENG: so that won't work either. 125UMI: yeah, that's not good.</Paragraph>
      <Paragraph position="2"> (4) 3RIC: I'm Ric M. and I am &lt;P&gt; what do I do? (whispering) software * yeah, I'm working for a software account 2  was recorded at a German University. Sin these dialogues, recorded for scenario design, speakers are assigned a new identity. In example (1), yeah is a positive answer to a proposal. In (2), it functions as a signal by means of which the speaker introduces a new topic, here it opens up the conversation, constituting the first utterance in the dialogue; in example (3), it serves to give feedback to the hearer, indicating perception and understanding and also basic agreement with the partner (Schegloff, 1982). Finally, in example (4), yeah occurs after a speech management problem, that is, after the speaker does not know how to continue.</Paragraph>
      <Paragraph position="3"> Other discourse particles may fulfill similar functions: in example (5), the discourse particle mmm functions as a signal of perception and understanding, and so does the first well in example (6) while it furthermore relates the current utterance to the previous. The second instance of well in this example marks a thematic break, concluding the topic concerning the partner's cold and returning to the previous attempt to find a possible date for a meeting. This discourse structuring function can also be found for the hesitation marker uh in example (7).</Paragraph>
      <Paragraph position="4">  16UMI: yeah, just one of the students about San Soto anyway.</Paragraph>
      <Paragraph position="5"> 17ENG: mram (6) 6UMI: I've got &lt;P&gt;  7ENG: well, you've got a very heavy cold, but well &lt;P&gt; wh/ wh/why is this week out of the question? (7) ll4UMI: yeah, this weekend's k k k completely chaotic, uh. let's see.</Paragraph>
      <Paragraph position="6"> Thus, the functions discourse particles may fulfill are as follows: they mark thematic breaks, thus making the macro-structural organisation of the dialogue transparent for the hearer; they signal the relatedness of the current utterance to the previous, they indicate whether the information transfer is successful, and whether the channel is still available. Finally they support the formulation process in case there are speech management problems. Thus, they may fulfill a large but not an unlimited number of functions.</Paragraph>
      <Paragraph position="7">  Furthermore, several of these functions have been found to co-occur in stable feature bundles. For instance, in example (6), the first instance of well displays perception and understanding of what the partner has said and furthermore indicates that the speaker is going to add something to the same topic. The same information is conveyed by yeah in example (3). This implies that there are not only co-occurrence relations between different functional features but also that a general inventory of pragmatic interpretation of discourse particles can be proposed, such that the same pragmatic function can be fulfilled by different discourse particles. The linguistic model presented here consequently assumes that each discourse particle can be assigned a context-specific interpretation (StenstrSm, !994) which belongs to a limited inventory and which can be seen as a bundle of functional features which may be fulfilled by several different discourse particles. Additional information is carried by the lexemes themselves (Fischer, 1996), for instance, in the case of yes agreement, and rejection in the case of no.</Paragraph>
      <Paragraph position="8"> The descriptive inventory, which was arrived at by means of several hypothesis-testcycles on the basis of four large corpora (Fischer andJohanntokrax, 1995; Schmitz and Fischer, 1995), consists of the following items: * take-up: signals perception and understanding regarding the previous utterance and takes up the current topic;  * backchannel: gives feedback and supports the other's turn; * frame: marks a break in the thematic structure; * repair marker: signals problems in the formulation process; * answer: signals agreement on, or rejection of, a proposition; * check: signals the hearer that the  speaker would like to get positive feedback; null * modal: refers to the common ground or the information 'on hand;' establishes a relation of coherence between the utterance and a pragmatic pretext.</Paragraph>
      <Paragraph position="9"> In how far can these functions now be used for automatic speech processing? On the one hand, the most trivial point is the ambiguity of certain discourse particles concerning propositional and other information, for example German ja, nein, doch or nicht, or English yes or no. Information on this ambiguity is essential for automatic speech processing systems since if it is not resolved, contradictory or just wrong information will be inferred. Consider the following proposal for an appointment in example (8).</Paragraph>
      <Paragraph position="10">  (8) geht es nicht urn 13 Uhr.C/ (isn't it possible at 1 p.m.?)  In this example the discourse particle nicht does not contribute to the propositional information; however, it does so in the following example: (9) es geht nicht urn I3 Uhr.</Paragraph>
      <Paragraph position="11"> (it isn't possible at 1 p.m..) The examples demonstrate the need for the disambiguation of discourse particle occurrences since the recognition of keywords such as nicht would lead to wrong conclusions regarding the first example; however, disregarding it would obscure the proposition in the second example.</Paragraph>
      <Paragraph position="12"> On the other hand, if the different functions of discourse particles can be automatically determined, they can also support further semantic and pragmatic analyses concerning the information structure and the macro-structural organisation of a dialogue, segmentation, and the recognition of repair sequences in the same way as they are taken as signals by human speech processors. Automatic disambiguation of discourse particles is therefore essential to their further employment in automatic speech processing systems.</Paragraph>
    </Section>
  </Section>
  <Section position="4" start_page="109" end_page="111" type="metho">
    <SectionTitle>
3 Automatic Disambiguation
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="109" end_page="109" type="sub_section">
      <SectionTitle>
of Discourse Particles
</SectionTitle>
      <Paragraph position="0"> To account for the multifunctionality of discourse particles, two strategies are combined: the analysis of the syntactic position of a discourse particle and the role of the pragmatic environment the discourse particle occurs in.</Paragraph>
    </Section>
    <Section position="2" start_page="109" end_page="109" type="sub_section">
      <SectionTitle>
3.1 The Influence of the Position in
Turn and Utterance
</SectionTitle>
      <Paragraph position="0"> Positions in turn and utterance seem to be good candidates for determining the interpretation of German discourse particles; for instance, if ja occurs utterance finally, it is likely to be used in a checking function, for example: (10) elf Vhr sagten Sie, ja? (FRS001) (you said l 1 a.m., right?) Thus, different functions can be assigned to certain contextual positions of ja; as we have seen, turn-final ja usually functions to elicit feedback from the hearer; utterancemedial ja, if the utterance is not interrupted and a new utterance is begun, serves as a modal particle, otherwise as a repair marker. The largest group is constituted by turn- and utterance-initial discourse particles; these serve either to signal perception and understanding and that one is going to add something to the same subject matter, or as answer signals. The latter may however also occur as a complete turn or turn-finally, so that other types of information are necessary to identity' the function of ja as an answer signal carrying propositional content than the position in turn and utterance only: Nevertheless, the syntactic position in which a discourse particle occurs provides a valuable source for the automatic disambiguation of discourse particles, as will be shown by means of the following experiments.</Paragraph>
      <Paragraph position="1"> First of all, 150 examples of the especially multifunctional German discourse particle ja were analyzed regarding their surface and pragmatic properties. Secondly, to each discourse particle a context-specific interpretation from the inventory proposed in the previous section could be assiglJed. By means of an artificial neural network classificator (Scheler and Fischer, 1997), it was determined whether the eight classes described in the previous section can be assigned automatically on the basis of a selection of structural and pragmatic features. In particular, training a feedforward-backpropagation network simulated by SNNS p.d. software with ca. 100 examples of German ja, and testing on ca. 50 new examples can show firstly whether the classification proposed is sufficiently contrastive and thus learnable. Furthermore it is determined of which influence on the correct classification of the new exarnples particular properties are such as the syntactic position or the respective intonation contour.</Paragraph>
      <Paragraph position="2"> Good results in the experiments employing the whole feature set indicate that the classes are learnable and that the classification can be replicated by an automatic classifter (Scheler and Fischer, 1997). Concerning the role of particular surface features, ca. 55% (standard deviation 4.38) of the context-specific interpretations can be determined only on the basis of surface properties such as syntactic position, combination with other discourse particles, intonation contour, or tim left and right context. The position in turn and utterance alone produces similar results. On the basis of the intonation contours alone, ca. fifty percent can be assigned correctly (however, the standard deviation here is quite high). Position and prosody together reach 57.7% (standard deviation 4.31). The other surface properties therefore only provide redundant information compared to these two types, adding them for training and testing the network does not increase the assignment rate. So it seems promising to regard aspects of the surface realization as a valuable source for the automatic disambiguation of the functions of discourse particles. This information is exploited in the implementation of a speech processing system in the appointment scheduling domain described in section 4 especially regarding the position with respect to the utterance since it is the information most easily to obtain for an automatic</Paragraph>
      <Paragraph position="4"/>
      <Paragraph position="6"> speech processing system.</Paragraph>
      <Paragraph position="7"> The first type of information which the automatic speech processing system is going to rely on for the automatic disambiguation of discourse particles is thus the prototypical syntactic position for each discourse function. null</Paragraph>
    </Section>
    <Section position="3" start_page="109" end_page="111" type="sub_section">
      <SectionTitle>
3.2 The Role of a Dialogue Model
</SectionTitle>
      <Paragraph position="0"> As a second major source of information for the automatic disambiguation of discourse particles, a significant correlation between certain dialogue acts, domain-specific speech acts (Schmitz and Quantz, 1995), and certain discourse functions, the context-specific readings of discourse particles, was determined. For instance, in the informal task-oriented human-to-human dialogues, 40.43% (standard deviation 7.9) of the discourse functions of ja could be assigned correctly by the artificial neural network classificator only on the basis of the information which the preceding and which the following dialogue act was, although in this corpus the number of dialogue acts assigned is quite high, and the relationship between them is highly variable. Consequently, the reduction of complexity caused by the specialization on one domain in task-oriented settings can be exploited here by postulating a limited inventory of speech acts occurring in the corpus under consideration. A dialogue model which represents the possible relations between dialogue acts in a flow chart can thus make predictions about the following dialogue act and also about a probable function of the respective discourse particle. Thus, information for each discourse particle about its occurrence in certain dialogue acts may support its automatic disambiguation. In human-computer interaction, for example, this sequential relationship between dialogue acts may be even more straightforward, and the domain may be furher restricted.</Paragraph>
      <Paragraph position="1"> For analyses of the appointment scheduling domain, a small corpus of 30 dialogues with altogether 145 turns was gathered in a hun~an-to-machine scenario to arrive at statistical correlations between functional information concerning discourse particles and tile dialogue acts in which they occur. On the basis of these instances of spoken interaction, a dialogue model was constituted (see Fig. 1). Tile descriptive categories were based on those proposed for tile Verbmobil corpus (Alexandersson et al., 1997). By means of these dialogue acts, it can be partly determined what the fun,tion of a respective discourse particle may be in the particular setting; for instance, it is much more likely that ja is an answer particle if it occurs in an utterance after the system has asked for the confirmation of a date (as in example  11) rather than an uptaking signal when it is used at the beginning of a dialogue between system S and user U (as in example  (Please confirm this appointment.) U: ja, dieser Termin ist richtig.</Paragraph>
      <Paragraph position="2"> (10_2_02) (yes, this is the right date.) (12) U: ja, guten Tag, ich hPStte gern einen Termin. (6_1_00) (yes, hello, I'd like to schedule an appointment.) null In the linguistic model, information on the correlation between dialogue acts and certain functions of discourse particles as well as syntactic information are combined. The following section describes the implementation of these types of information.</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="111" end_page="111" type="metho">
    <SectionTitle>
4 Implementation
</SectionTitle>
    <Paragraph position="0"> The model described above was implemented and tested in a system for appointment scheduling (Brandt-Pook et al., 1996). For the representation of linguistic knowledge, this system uses the semantic network language ERNEST (I(ummert et al., 1993). The linguistic knowledge of this system is modelled hierarchically on five levels of abstraction: The word hypothesis level establishes the interface to the speech recognition system; the syntactic, semantic and the pragmatic level declaratively represent the interaction of lexical and other knowledge; and the dialogue level manages dialogue strategy and dialogue memory. The concepts at the different levels are connected by means of abstraction links. Within a level, has-part and is-a links are defined.</Paragraph>
    <Paragraph position="1"> Each occurrence of a discourse particle is interpreted bottom-up. Thus, a wordhypothesis is created so as to identify the relevant word form; connected to this is a concept PARTICLE, for which the utterance position for the respective discourse particle is identified. On the dialogue level the interpretation of a discourse particle occurrence and its functional role is completed.</Paragraph>
    <Paragraph position="2"> The attribute att_disc_particle realizes rules that combine the different types of information which participate in the model for the pragmatic interpretation of discourse particles. Thus, on the one hand, the position in turn and utterance is treated as a condition on the interpretation of a discourse particle, on the other, due to the fact that the previous system-output influences the next user utterance, the dialogue act of the following utterance and consequently the function of a respective discourse particle can be predicted. For instance, a system request for an explicit confirmation predicts an answer signal by the user, as implemented in the rule concerning the word form ja (see Fig. 2). With this information, an utterance  answer signal or as a discourse particle with certain pragmatic functions.</Paragraph>
  </Section>
class="xml-element"></Paper>